A 1.4 KB CSV file by Kun Xiang, last updated May 22, 2026, quantifies the reliability gap in plant disease classifiers moving from controlled to field imagery. It benchmarks standard mitigation techniques like temperature scaling and selective prediction, showing a 67.7-percentage-point accuracy collapse under domain shift. The analysis includes quantitative image-level shift metrics and performance results for models like ResNet-50 and DINOv2 across datasets PlantVillage and PlantDoc.
Use Cases
- Benchmarking confidence calibration techniques based on reported metrics like calibrated ECE and selective risk.
- Evaluating domain adaptation methods for agricultural vision based on results for adaptive batch normalization and feature moment matching.
- Comparing model backbone performance on out-of-distribution detection based on the reported AUROC scores for DINOv2 and ResNet-50.
- Analyzing the impact of image-level domain shift based on reported effect sizes for saturation, border edge density, and foreground occupancy.
Strengths
- Includes specific, reproducible performance metrics such as a 67.7-percentage-point accuracy drop and a calibrated ECE of 0.3645.
- Benchmarks multiple standard mitigation techniques and model architectures, providing comparative results.
- Quantifies domain shift with concrete effect sizes (e.g., d=3.90 for saturation differences).
Limitations
- Row count is unknown, which may limit suitability assessment.
- Column-level documentation is absent; field semantics must be inferred after download.
- The dataset is very small (1.4 KB), indicating it likely contains summary statistics rather than raw image data.
Provenance
- Source
- figshare
- Collection Method
- Models trained on PlantVillage were evaluated on PlantDoc under a parent-image-aware split protocol, with a suite of standard mitigation techniques applied.
- Freshness
- Last updated 2026-05-22 06:11:02; freshness should be verified.